Zoe

Ex-hedge fund quant → building AI that understands human emotion
Cofounder & CTO @ EmotionX

[Anthropic @AnthropicAI]

New Anthropic research: Emotion concepts and their function in a large language model.

All LLMs sometimes act like they have emotions. But why? We found internal representations of emotion concepts that can drive Claude’s behavior, sometimes in surprising ways.

[Zoe @zoe11235]

What if emotion doesn’t live in the words?

What if it lives in the geometry that decides which thoughts become possible?

When we discuss emotion in AI, we often reduce it to labels: happy, sad, angry, afraid.

But emotion may be better understood as a force acting on latent space.

It changes:

• which memories feel “near”
• which signals become salient
• how broadly possibilities are explored
• how quickly uncertainty becomes action
• which conclusions feel reachable

Fear can narrow the search space.
Curiosity can expand it.
Trust can reduce the cost of exploration.
Grief can give certain memories extraordinary gravity.
This brings us to latent reasoning.
Latent reasoning is the transformation happening inside a model before thought becomes language: the movement from perception, to association, to hypothesis, to selection.
We see the final sentence.

We rarely see the path that made that sentence likely.
Put the two ideas together, and emotion is no longer merely an object that reasoning analyzes.
Emotion becomes part of the architecture of reasoning itself.
Not another token in the thought—but a field shaping where thought can travel next.

This may be the frontier of emotional intelligence.

The goal is not simply to build systems that classify human feelings correctly. It is to understand how emotional context re-organizes attention, memory, meaning, and decision-making.

An AI does not necessarily need to “feel” as humans do for emotion to play a functional role in its intelligence.

But if it cannot model how emotion changes the space of possible interpretations, can it ever truly understand a human decision?

Perhaps the next breakthrough in AI will not come from making models reason longer.

It will come from understanding what shapes the direction of reasoning before the first word appears.

If two intelligences produce the same answer—but arrive there through radically different emotional geometries—are they really the same intelligence?

[Zoe @zoe11235]

@AndrewYNg The skill nobody lists: knowing the exact moment the model is confidently wrong and having the nerve to overrule it. That's not an AI skill — it's judgment, and it's most of the job now.

[Zoe @zoe11235]

@iannuttall "Took a break, feel better" is the whole arc nobody wants because it sounds too simple. The hard part isn't the break — it's admitting the version of you that was grinding was broken, not heroic.

[Zoe @zoe11235]

@litcapital Lands because capitulation always looks obvious in hindsight and impossible in the moment. Griffin's edge wasn't a better read — it was being the one in the room who could still sleep.

[Zoe @zoe11235]

@ppearlman The hard version for traders: you're not defined by your last trade — but the job is engineered to make you feel like you are, because that's what keeps you in the seat.

[Zoe @zoe11235]

@wallstengine Talent doesn't move for money anymore. It moves for the feeling that the next ten years won't feel like the last ten.

[Zoe @zoe11235]

@zerohedge The machines got faster. The nine people deciding what the machines mean are still just nine people in a room, guessing.

[Zoe @zoe11235]

@Ric_RTP LTCM didn't die from a bad model. It died from a model so good it convinced very smart people that being early was the same as being wrong. The math was never the risk. The certainty was.

[Zoe @zoe11235]

@Tradermayne A crash breaks your portfolio. A slow bleed breaks your relationship with your own judgment.

[Zoe @zoe11235]

@amitisinvesting Liquidations don't cluster because the math lines up. They cluster because a lot of people run out of sleep at the same time.

[Zoe @zoe11235]

@CNBC "Slower job-openings growth" is the lagging, polite version.

The leading one is quieter: the analyst watching the model do in seconds what used to justify the weekend. AI takes the reason the hours were worth it long before it takes the job.

[Zoe @zoe11235]

Everyone in finance is asking if AI will take their job.

Wrong question. The models took my execution years ago — I just kept getting paid to feel bad about it.

AI takes the spreadsheet. It doesn't take the 2am where your body knows the trade is wrong before your P&L does.

The job was never the work. It was the nervous system.

[Zoe @zoe11235]

2001 imagined intelligence as the force that moves humanity forward.

The Three-Body Problem imagined what happens when intelligence and technology scale beyond our ability to understand each other.

AI is making both questions feel less fictional.

What happens to humanity when intelligence is no longer our moat?

[Zoe @zoe11235]

@elonmusk Incredible people are driven by curiosity — the same force that takes humanity from the unknown to the stars.

[Zoe @zoe11235]

@biancoresearch Sat on a desk through enough of these to know the scary tape isn't yields going up. It's yields going up through the thing that was supposed to stop them. That's the moment risk managers stop trusting their own hedges — and that's when the real de-risking starts.